Brendan Gregg on Linux Performance
# reset collision, drop, etc counters reported by ifconfig and ethtool -S ethtool -i enp94s0f1 # determine driver used (if not build into kernel) ifconfig enp94s0f1 down modprobe -r ixgbe # intel 10G NIC modprobe ixgbe ifconfig enp94s0f1 up ethtool -S enp94s0f1 | egrep -i 'drop|col|err' # look for drop or collision, no report on pause frames ethtool -S enp216s0f1 | grep -v ': 0$' | more # look for anything with non zero counter ethtool -G enp180s0f0 rx 8192 tx 8192 ethtool -g enp180s0f0 # lower case for query, upper case for set ethtool -C enp180s0f0 rx-usecs 100 ethtool -c enp180s0f0 mlx5_core # mellanox 100G Ethernet (and IB?) ss -tulp # check socket status, see which port is open by what process. add -n to show port by number
vmstat 1 5 procs -----------memory---------- ---swap-- -----io---- -system-- ------cpu----- r b swpd free buff cache si so bi bo in cs us sy id wa st 15 0 0 7207116 31284 1328320 0 0 0 0 24 12 0 0 99 0 0 12 0 0 7207084 31284 1328320 0 0 0 0 88618 152257 44 1 55 0 0 12 0 0 8030280 31284 1328224 0 0 0 12 88673 152364 44 1 55 0 0 13 0 0 7210160 31284 1328224 0 0 0 0 88686 152315 43 1 55 0 0 12 0 0 7210284 31284 1328224 0 0 0 0 88683 152378 44 1 55 0 0
Goal is to determine how much memory to request for an hpc batch job. eg, what value to give "qsub -l m_mem_free".
Ideally, there is a memory calculation equivalent to "time -p CMD".
But, AFAIK, there isn't.
So, these things will help finding a decent number for memory request.
- ps aux -q PID, look at VSZ (in kB) column of the running process.
- cat /proc/PID/status, and look for VmPeak and VmSize (value should be similar to ps aux above) .
- run top/htop and find the process and see how much memory is being reported under VIRT .
- qacct -j JOBID, if accounting is enabled.
The "mem" entry is in "GBytes CPU seconds" :(, but divide that by ru_wallclock and it give some idea.
The "maxvmem" is largest amount memory the job used.
- perf_events don't seems to offer anything useful for this :(
- TCMalloc ?
- Valgrind ?
SATA 150 has 150 MB/s bandwidth SATA II has 300 MB/s SAS is 6 Gbps = 768 MB/s PCIe 3.1 x4 NVMe = ... NVMe. Replaces AHCI. Needed support from OS/kernel, but widely available ~2015, show up as /dev/nvme*. NVMe drives are often in NAND. Avail format: M.2, U.2, AIC (Add-In card, ie pci card) M.2 = replaces mSATA * typically on mboard, internal to machine, 3.3V only * NOT hot swappable--often on laptop, but number of servers use them too :-\ * typically bare chips, gumstick format U.2 (SFF-8639) * The U.2 connector is mechanically identical to the SATA Express device plug, but provides four PCI Express lanes through a different usage of available pins (wikipedia) * Some nvme drive support sata as well for backward compatibility, but at reduced bandwidth. * ie does not use SATA controller, straight to the PCIe bus * support hot swap, 3.3V or 12V. * typically inside an enclosure, mounted on a hot swap disk tray, like traditional 2.5" HDD AIC * Add-In card format. ie pci card. * back then hard some AIC card better performing than U.2 (not sure why, cuz U.2 should be PCIe x4 too). * Some AIC is just carrier for 2x U.2 drives. b = bits (network, serial protocol)) B = Byte (storage, at least traditionally "parrallel", whole byte based protocol) Ti = Si unit, ie multiple of 1000, not 1024. 128KB 4KB Latency Power Seq read/write Rnd r/w read/write active/idle MB/s IOPS 1000 microsec w --------------------- ------------- ----------- -------- ------ ------- Intel SSD S4510 560/510 97 /36 Intel SSD S4610 560/510 97 /51 Intel SSD P4510 1TB 2850/1100 465 /70 77/18 10/5 64-Layer 3D TLC NAND Intel SSD P4510 2TB 3200/2000 637 /81.5 12/5 Intel SSD P4510 8TB 3200/3000 641 /134.5 15/5 Intel SSD Pro 7600p M.2 3200/1625 340k/275k 3D2 TLC 512 GB to 2 TB version Intel SSD Pro 7600p M.2 1640/650 105k/160k 128 GB version Micron 9100 Max Much older, no longer avail? Micron 9200 Max U2 3430/2400 800/270 Micron 9200 Max AIC 4600/3800 1000/270 Micron 9300 Pro U2 3.8T 3500/3100 835/105 86/11 Micron 9300 Pro U2 7.6T 3500/3500 850/145 Micron 9300 Max U2 3.2T 3500/3100 835/210 Micron 9300 Max U2 6.4T 3500/3500 850/310 TLC = Tri-Level Cell. 3-bit addressing, maybe 32 or newer use 64 layers. Larger capacity drive sometime has better performance, some as good as 2x! Micron typically best. Intel are decent. Samsung typically worse, consumer oriented device. Performance/Lifespan is probably 1 year. Micron Pro are Sequential Read optimized, Max are good for random access. Max reserve more cell to replaced expected failure of random access workflow (ie with many write/erase) Seq r/w Rnd r/w MB/s MB/s IOPS --------------------- ------- --------- ---- Samsung HD501LJ 7200rpm 80-90 WD Caviar Green 5400rpm 223 WD Red 5400rpm 368 Seagate Barracu 7200rpm 410 ? 120 is what NetApp use for calculations). For performance rule of thumb, estimate 7200 RPM drive to give about 100 to 150 MB/s. This would be combination of sequential and random IO which would mean about 1MB per IO-operation for drive doing 100-150 IOPS.Ref:
numactl -H # should be in FLAT mode (ie, at least two NUMA domain, so that MCDRAM vs off-socket DDR4 memory can be addressed separately) # IF in cache mode, effect of in-socket memory caching cannot be removed from stream test export OMP_NUM_THREADS=64; export KMP_AFFINITY=scatter; numactl -m 0 ./stream
cstate and pstate, which govern powering down subsystem and voltage/frequency, respectively, helps save power but can cause unexpected latency in PCI bus and thus Infiniband performance.
Things to tweak:
1. remove your current kernel args "processor.max_cstate=1 intel_idle.max_cstate=0"
2. add the new kernel args "intel_pstate=disable"
wwsh -y provision set n0143.savio3 --kargs="nopti intel_pstate=disable processor.max_cstate=1 intel_idle.max_cstate=0"
# didn't help with colefax gpu
wwsh -y provision set n0143.savio3 --kargs="nopti processor.max_cstate=1 intel_idle.max_cstate=1"
# osu_latency, osu_bw still bad , even after cpupower -g performance
(very old ver of wwulf may only keep first kargs param.)
cat /proc/cmdline after machine boot to see kernel arg actually used to boot.
ref: https://access.redhat.com/solutions/202743
https://www.kernel.org/doc/html/latest/admin-guide/kernel-parameters.html
intel_idle.max_cstate= [KNL,HW,ACPI,X86]
0 disables intel_idle and fall back on acpi_idle (ie /sys/devices/system/cpu/cpuidle/current_driver=none)
1 to 9 specify maximum depth of C-state.
https://www.thomas-krenn.com/en/wiki/Processor_P-states_and_C-states
C0 – Active Mode: Code is executed, in this state the P-States (see above) are also relevant.
C1 – Auto Halt
C1E – Auto halt, low frequency, low voltage
C2 – Temporary state before C3. Memory path open
C3 – L1/L2 caches flush, clocks off
C6 – Save core states before shutdown and PLL off
C7 – C6 + LLC may be flushed
C8 – C7 + LLC must be flushed
disabling pstate seems to have significant impact on infiniband latency (osu_latency), but then keep cpu running at normal clock rate and not slowing down.
cstate doesn't seems to have impact on osu_latency. ie left /sys/module/intel_idle/parameters/max_cstate = 9
pdsh -w n0[143-145].savio3 "cat /proc/cpuinfo | grep 'cpu MHz' | head -1"
cat /sys/module/intel_idle/parameters/max_cstate # 9 is default. 0=disabled
kargs not likely useful. cmd should work, but need them installed and invoked on boot...
Intel Cascadelake, kargs="quiet nopti" ==> cpupower driver: intel_pstate # n0229.sav3
Intel Cascadelake, kargs="nopti intel_pstate=disable" ==> cpupower driver: no or unknown cpufreq driver is active on this CPU # n0138.sav3
AMD Epyc Rome, kargs="quiet nopti" ==> cpupower driver: acpi-cpufreq # n0209.sav3 Epyc 7302
cpupower frequency-set -g performance
# 800 or 1200 is low
# 2600 would be more performant, though it does use more power.
cat /sys/devices/system/cpu/cpu*/cpufreq/cpuinfo_min_freq
cat /sys/devices/system/cpu/cpu*/cpufreq/cpuinfo_max_freq
cat /sys/devices/system/cpu/cpu*/cpufreq/cpuinfo_cur_freq
(seems not avail if psate is disabled, ie, when using acpi driver instead of intel_pstate driver (?)
but cat /proc/cpuinfo | grep MHz is reflective of cpuinfo_cur_freq )
turbostat
kargs intel_pstate=disable would (mostly*) make cpu run at a single frequency.
lscpu would then reports a single fixed frequency instead of Max/Min.
(mostly*) = Most of the time. BIOS pstate setting on Tyan mboard w/ casdelake cpu also need to be disabled.
Power utilization tends to go up
(for Linux kernel 3.x, the default is to use acpi driver when intel_pstate=disable is ABSENT)
Tests:
mpirun -np 2 -H 10.0.7.144,10.0.7.145 -report-bindings -bind-to core osu_latency # need to be able to ssh to node to run orte
# latency below 1.20 us would be good; ~2.2 is common. 3+ us starts to be bad. 5.6+ would be bad. 10+ would be horrible
# PCI switch would add latency of ~0.3 us.
# cpu binding to PCI slot of IB card is needed for optimal performance, ~0.4 us.
# ethernet latency is in ms (? maybe ~100 us).
# OSU_latency maybe subject to (mpi) timing bug. at least running linpack with mpi may get affected depending on clock frequency setup.
mpirun -np 2 -H 10.0.7.144,10.0.7.145 -report-bindings -bind-to core osu_bw
# Size Bandwidth (MB/s) bad eg:
# good eg: bad eg:
1 2.67 1.55
2 5.34 3.51
4 12.00 7.32
...
32768 9956.58 7977.42
65536 11726.98 9445.02 # good BW plateau much earlier and at higher value
...
2097152 12403.11 10018.76
4194304 12411.50 10024.22
date ; time -p mpirun -np 2 -H 10.0.7.195,10.0.7.196 -report-bindings --bind-to core --cpu-set -0-15 osu_latency; date
[n0195.savio3:42894] MCW rank 0 bound to socket 0[core 0[hwt 0]]: [B/././././././././././././././././././.][./././././././././././././././././././.]
[n0196.savio3:43753] MCW rank 1 bound to socket 0[core 0[hwt 0]]: [B/././././././././././././././././././.][./././././././././././././././././././.]
# OSU MPI Latency Test v5.4.0
# Size Latency (us)
0 2.46
1 2.54
2 2.32
4 2.32
...
2097152 327.63
4194304 648.33
real 4.07
user 0.10
sys 0.12
3a. Bios Changes
I went to the bios of two asus nodes (n0137.savio3, n0138) and made a number of changes.
The latency, measured only between two nodes, is ~3.2 us. Better than the 4 us originally, but still not the 1.2 us we are getting from the supermicro nodes.
disable hyperthreading (was enabled)
Intel TXT left as disabled
VMX left as enabled
SMX left as disabled
there were lot of profile optimization for various benchmark, like MP Linpack, spec cpu, HammerDB, STREAM Triad, etc.
mostly, think it change Engine Boost
core optimizer is always disabled for many benchmark/workflow i selected
end up picking by workload (recommended to operate at 25C or below, so not sure if we should leave this on?)
picked HPC
(other options: Latency optimized, Peak Frequency, power efficient)
change
Engine Boost -> Level3(Max) (was disabled
core optimizer is left as disabled
P-state: (left all as default)
boot performance: max
energy efficient turbo: enable
turbo mode: enable
c-state
not entirely obvious
disabled CPU C6 report ##
disabled C1E
package c state control
changed from auto to C0/C1 state ##
left sw controlled T-states as disabled (default)
trusted computing
Changed to "disabled" security device
none was found to begin with anyway
ACPI settings
Disabled hibernation
3b. A second BIOS change that I tested, but no real improvements either:
Socket Config
CPU - Adv PM Tuning
Energy perf BIAS
Workload config
- first test above had left as Balanced
- 2nd test changed to I/O sensitive
Ref:
7xx1 = 1st Gen Epyc Naples 7xx2 = 2nd Gen Epyc Rome ca 2019 7xx3 = 3rd Gen Epyc Milan ca 2020.11
C-states should be left enabled in BIOS.
C2 can be disabled for IB via cli.
- C0: active, the active frequency while running an application/job.
- C1: idle
- C2: idle and power-gated.
This is a deeper sleep state and will have a greater latency in getting to C0
compared to when it idles in C1.
For a HPC cluster with a high performance low latency interconnect such as Mellanox
disable the C2 idle state. Assuming a dual socket 2x32 core system
cpupower -c 0-63 idle-set -d 2
• Set the CPU governor to ‘performance’:
cpupower frequency-set -g performance
Use
hwloc-ls to find out which cpu id is closest to the IB
then use
numactl -C $CPUID osu_bibw
to bind the MPI process (eg osu_* microbenchmark binary) to the cpuid closest to the IB
for lowest latency.
Virtual Memory reclaim (good to run in epilog to free up cache)
/proc/sys/vm/zone_reclaim_mode sysctl vm.zone_reclaim_mode=1 # default is 0 sysctl vm.swappiness=10 # default is 60, lower number use less swap sync echo 1 > /proc/sys/vm/drop_caches [frees page-cache] echo 2 > /proc/sys/vm/drop_caches [frees slab objects e.g. dentries, inodes] echo 3 > /proc/sys/vm/drop_caches [cleans page-cache and slab objects]Max performance is to have most NUMA that can be supported. Each socket can do 1,2,4 NUMA (zone?). For NUMA=4 req all DIMM to be populated? But if use NUMA=2 or 1, how to run HPL with OpenMP?
docker login nvcr.io # Username: $oauthtoken # Password: cWR2az... CONT='nvcr.io/nvidia/hpc-benchmarks:21.4-hpl' # linpack CONTcg='nvcr.io/nvidia/hpc-benchmarks:21.4-hpcg' # congutate gradient CONT_i='registry:443/tin/hpc-benchmarks:21.4-hpl' # internally hoster registry docker pull $CONT docker tag nvcr.io/nvidia/hpc-benchmarks:21.4-hpl registry:443/tin/hpc-benchmarks:21.4-hpl docker image push registry:443/tin/hpc-benchmarks:21.4-hpl docker run -v /global/home/users/tin:/mnt --gpus all $CONT_i \ mpirun --mca btl smcuda,self -x UCX_TLS=sm,cuda,cuda_copy,cuda_ipc --bind-to none -np 2 \ hpl.sh --cpu-affinity "0-7:120-127" --gpu-affinity "0:1" --cpu-cores-per-rank 8 --dat /mnt/hpl_dat/HPL-a10-Ns=20k.dat #HPL-a10.dat docker run -v /global/home/users/tin:/mnt --gpus all registry:443/tin/hpc-benchmarks:21.4-hpl mpirun --mca btl smcuda,self -x UCX_TLS=sm,cuda,cuda_copy,cuda_ipc --bind-to none -np 2 hpl.sh --xhpl-ai --cpu-affinity "0-7:120-127" --gpu-affinity "0:1" --cpu-cores-per-rank 8 --dat /mnt/hpl_dat --dat /mnt/hpl_dat/HPL-a10-Ns72k_1x2.dat
ps -e -o pid,vsz,ruser,comm= | sort -n -k 2
# show all process, sorted by virtual memory size (Kolumn 2)
ps -efl | sort -n -k 10
# show all process with lots of details, sorted by SZ (Kolumn 10)
ps fields:
rss RSS resident set size, the non-swapped physical memory that a task has used (in kiloBytes).
(alias rssize, rsz).
size SZ approximate amount of swap space that would be required if the process were to dirty all
writable pages and then be swapped out. This number is very rough!
sz SZ size in physical pages of the core image of the process. This includes text, data, and
stack space. Device mappings are currently excluded; this is subject to change. See vsz
and rss.
vsz VSZ virtual memory size of the process in KiB (1024-byte units). Device mappings are
currently excluded; this is subject to change. (alias vsize).
netstat -ta show current intenet services/connections
-a : show (a)ll (include listening port process)
-n : ip (n)umber only (no dns lookup)
-r : (r)outing table (change with route cmd)
-i : show stat for diff nic (i)nterfaces
-k ce0 : lot of interface specific info, ce NIC will have duplex stat.
netstat -tulpn # list LISTEN port and process
netstat -pan # list process and ports it has open
netstat -p : print ip to mac address table known to host (solaris?)
netstat -k : print lot of kernel stat, among it hme0 is for the sun's build in
happy meal ethernet nic (see sunsolve infodoc 17416 for explanation of these
undocumented stats, good for trobleshooting network latency, comare agaist cisco stat.)
netstat -s : show high level packet send/receive/fragment info
vmstat -a : all
-n :
-p : process owning port
iostat -xn 30 : check for disk activity, anything more than 5% busy and avg resp time > 30 ms is bad.
nfsstat
mpstat 10 : processor stats, repeat every 10 seconds
: In Solaris, it reports context switch, interrupt, mutex spin, xcal, etc
: see http://sunsite.uakom.sk/sunworldonline/swol-08-1998/swol-08-perf.html
mpstat -A = -u -I ALL -P ALL
-u = report cpu util
-I ALL = interrupt stat
-P ALL = all processor
cpustat : find out what cpu is doing...
truss -c -p PID : find number of system call and usr time for a process (sol, linux use strace)
lockstat sleep 5 : gather kernel lock stats during the sleep period (5 sec)
: solaris, run as root
protocol
sysmon
trapstat, thread list, kmastat, kmausers
GUDS
Guds is a script to gather performance stats for Solaris.
Sample usage is
./guds_2.4.5
./guds_2.4.5 -qX -H3 -s65040465
-qX is for quite mode
-H3 is for running it for 3 hours
-sNNNNN is the sun case number (info embeded in dirs created by guds to
store the files).
It collects lots of info in /var/tmp/CASEID/guds-DATE-TIME/...
May need lot of know how to analyze data.
Having a baseline when things is good and when there are
performance problems would help.
date; mkfile 1000m test; date # create a 1 GB file (filled with 0)
date; dd if=/dev/urandom of=test bs=1024 count=100000 # same, file has random data.
# encrypt /dev/zero, so that data will be kinda random.
# good speed for cpu with aes instruction
# eg could get 1.1 GB/s on a flash storage
dd if=<(openssl enc -aes-256-ctr -pass pass:"Hello" -nosalt < /dev/zero) of=/dev/null bs=1M count=500000 iflag=fullblock
net.core.wmem_max = 262144 net.core.wmem_default = 262144 net.core.rmem_max = 262144 net.core.rmem_default = 26214410GbE connected systems
net.core.wmem_max = 16777216 net.core.wmem_default = 524287 net.core.rmem_max = 16777216 net.core.rmem_default = 524287
yum install sysstat It install /etc/init.d/sysstat, run sadc to collect the stats. Data saved in /var/log/sa/saNN, where NN is the calendar day for which the stat is collected. Data is kept for 1 week. sadc -S SNMP is needed to collect history info for IP, TCP, UDP, ICMP. sadc -S IP6 is needed for v6 version of above. sar # shows most common stats sar -A # shows all recorded activities sar -r # memory and swap utilization (free vs used; 98% util for 4 days on end is okay) sar -R # memory rate stat (frmpg/s = freed mem page per sec. -ve = allocated pages) sar -S # swap utilization sar -w # swap rate sar -q # shows load avg, run queue length sar -u # combined cpu util sar -P ALL # per-cpu utilization sar -B # paging stats sar -b # io xfer rate # show report recorded on named file with start time of 8pm, end time of mid-nite, # at interval of 1800sec (30min) (default is 10min interval) sar -f /var/log/sa/sa22 -s 22:00:00 -e 23:59:00 -i 1800Live Network Reporting
sar -n KEYWORD 1 # live network report every 1 sec sar -n ALL 1-5 # report all network stats sar -n DEV 1 # check network dev utilization. reported in kBytes/s. sar -n DEV | grep enp1s0f0 # specific nic sar -n EDEV 1 # dev errors, eg rxfram reports frame alighment errors sar -n NFS,NFSD 1 # NFS client and server stats sar -n SOCK 1 # reports IP fragment in use (from mismatched MTU?) sar -n IP 1 # check network dev utilization. reported in kBytes/s. sar -n TCP,ETCP 1 # check TCP stats, and errors eg retransmit. active = outbound, passive = inbound
Basic SAR Setup (from HP-UX sys admin handbook and tooltips, p503): sar -o /tmp/sar.data 60 300 # run sar every 60 sec for 300 count, -o store info in file (bin) sar -u -f /var/adm/sa/saXX # read data from file (Solaris, XX = date number) sar -u -f /tmp/sar.data # read data from file (HP-UX) -u display cpu info (similar to iostat and vmstat) -b buffer cache activity, imp for oracle -d disk activity -q avg queue length (if run queue > num of cpu, will have to wait). -w swap info Setup SAR data collection for HP-UX (should also work for other platform): http://www.sarcheck.com/sarhowto.htm (Actually SarCheck.com, but cost money!) mkdir /var/adm/sa, then setup root crontab: #collect sar data # every 20 min 8-5, hourly outside normal work 0 * * * * /usr/lbin/sa/sa1 20,40 8-17 * * 1-5 /usr/lbin/sa/sa1 #reduce the sar data # generate pre-formated report focus for business hrs 5 18 * * * /usr/lbin/sa/sa2 -s 8:00 -e 18:01 -i 1200 -A # sample for SF + Minsk work hours (10 hours diff) 0 * * * * /usr/lbin/sa/sa1 15,30,45 0-8,10-19,23 * * 1-5 /usr/lbin/sa/sa1 05 21 * * * /usr/lbin/sa/sa2 -i 3600 -A # sa1 is data collection to /var/adm/sa/saXX # sa2 really produce condense version of report to /var/adm/sa/sarXX (sar vs sa) # filenames are reused every month. # use sar -A -f /var/adm/sa/saXX to get more detail report than std summary.
Solaris starts sadc in /etc/rc2.d/S21perf , a deamon to collect sar info. the script really run /usr/lib/sa/sadc /var/adm/sa/sa`date +%d`
AIX has preset entries in crontab for 'adm'. Check to ensure script exsit. sar logs are stored in /var/adm/sa
perf list # list events perf list | grep ext4 # look for fs related event perf stat CMD # count events from execution of "cmd", perf stat collect data till the CMD completes perf stat uptime # run the command uptime and show general cpu usage stat perf stat -p PID # hit ^C to see CPU usage stat of a running process collected while perf ran perf stat -a sleep 5 # system wide CPU usage, collect till CMD "sleep 5" completes perf stat -e 'ext4:*' ls # run the command ls and capture ext4 related events perf stat -e 'ext4:*' -a sleep 5 # collect ext4 events of all process, for 5 seconds, then exit and print output to screen perf record -F 99 -a -g -- sleep 10 # capture stack at freq of 99 Hz for 10 seconds # result captured to file perf.data perf record -F 99 -p PID # profile a running process, until ^C perf report # view perf record data stored in perf.data (interactive TUI) perf report --sort comm,dso perf report -n --stdio # output report to console perf annotate --stdio # annotated instructions w/ %, need debuginfo or will feel like reading assembly dump! perf script # print out all events as text # generate a "flamegraph" perf script > out.perf02 cat out.perf02 | ./stackcollapse-perf.pl | ./flamegraph.pl > out.svgflamegraph is stack trace across time (on x-axis), so to get idea what fn call is taking lot of time. The tool is avail from http://www.brendangregg.com/FlameGraphs/cpuflamegraphs.html, which will generate an image map on top of the SVG, allowing for interactive drill down!
# Sample CPU stack traces for the specified PID, at 99 Hertz, for 10 seconds: perf record -F 99 -p PID -g -- sleep 10 # Sample CPU stack traces for the entire system, at 99 Hertz, for 10 seconds: perf record -F 99 -ag -- sleep 10 # Sample CPU stack traces, once every 10,000 Level 1 data cache misses, for 5 s: perf record -e L1-dcache-load-misses -c 10000 -ag -- sleep 5 # Sample CPU stack traces, once every 100 last level cache misses, for 5 seconds: perf record -e LLC-load-misses -c 100 -ag -- sleep 5 # Sample on-CPU kernel instructions, for 5 seconds: perf record -e cycles:k -a -- sleep 5
setup env:
export PATH=$PATH:/opt/RICHPse/bin
export SEPATH=/opt/RICHPse/examples:/opt/RICHPse/toptool
interactive tools:
se zoom.se # gui, summary status for all components. Main Window.
se live_test.se # text version of zoom.se
se multimeter.se # gui, cpu, cache, vm and locks meter
se toptool.se # gui, just like top
se xload.se # gui, just like xload, show hostname :)
se infotool.se # gui, menu to lot of sys info (cpu, net, disk, etc)
se xit # gui wrap on text disk stat dump (xiostat.se)
se -DWIDE pea.se 10 # text, dump top like info to stdout every 10 sec
se disks.se # text, dump lot of disk usage info
se webtune.se # display current, min and max values for perf params
se virtual_adrain.se & # text, dump warning to stdout if perf problem found
# run cli in background, non permanent, only output to
# login screen; process end, all cleared.
-------------------------
# install:
# pkgrm RICHPse
# gunzip RICHPse.tar.gz
# tar xf RICHPse.tar
# pkgadd -d . RICHPse
# edit /opt/RICHPse/etc/se_defines, enable "disk nfs"
# alt, can just copy to network drive, and set PATH and SEPATH
# at least for the interactive tools above
# always run monitor:
/opt/RICHPse/etc/init.d/vader start # init.d script to start vader
se /opt/RICHPse/examples/vader.se # the "Virtual Adrian Daemon",
# start on host to be monitored
se /opt/RICHPse/examples/darth.se -h remotehost # gui, start on client.
# This gui is the front end of the bg monitor
#!/bin/sh # setoolkit-install.sh # quick script to setup and start se toolkit cd /mnt/sa/share/software/SEtoolkit pkgadd -d . RICHPse.331 (cd /opt/RICHPse/etc; tar cf - *.d) | (cd /etc ; tar xvf - ) # /etc/init.d/mon_cm start /etc/init.d/monlog start /etc/init.d/percol start /etc/init.d/va_monitor start /etc/init.d/vader start
lockstat sleep 5 : gather kernel lock stats during the sleep period (5 sec) : solaris, run as root
stress --cpu 60 -t 60s ## 152 W stress --cpu 64 -t 60s ## 160 W stress --cpu 200 -t 60s ## 160 W stress --vm 60 -t 60s ## 216W on knl savio2 # idle = 96W stress --vm 64 -t 180 ## 240W stress --io 60 -t 60s ## 144W stress --hdd 60 -t 60s ## 112W stress --io 4 --hdd 2 --vm 8 --cpu 50 -t 60s ## 160W stress --io 4 --hdd 2 --vm 50 --cpu 50 -t 60s ## 240W stress --io 6 --hdd 2 --vm 64 --cpu 64 -t 180s ## 240W # max power usage when cpu match core counts *** stress --vm 128 --cpu 128 -t 120s ## 232W #didnt crash :P stress --vm 128 --cpu 32 -t 120s ## 224W stress --io 32 --vm 128 --cpu 32 -t 120s stress --io 256 --hdd 2 --vm 256 --cpu 64 -t 120 ## 216W #didnt crash :P stress --io 6 --hdd 2 --vm 64 -t 120 ## 240W #dont need to hog cpu, vm will take care of it, io was matched to number of memory channel ## actually, just --vm matching number of cores pretty much reach peak power, --io, --hd don't add to power load once it is peaked. ## yet on AMD having both --vm and --cpu flag drive energy usage up. mprime -t got to similar level. ## linkpack likely only used 208W singularity exec /global/scratch/tin/singularity-repo/perf_tools_latest.sif stress --io 6 --hdd 2 --vm 64 -t 120 # each number after the test is number of worker # io test and virtual memory test automatically trigger substantial load on cpu already # -t 60s = limit test to 60 sec while [[ 1 == 1 ]] ; do ipmitool sensor | grep -i watt ; sleep 5; done while [[ 1 == 1 ]] ; do ipmitool sensor|egrep -i watt\|amp; lscpu|grep MHz; sleep 5; done singularity exec /global/scratch/tin/singularity-repo/perf_tools_latest.sif stress --io 6 --hdd 2 --vm 64 -t 120 ## 406W on savio3 bigmem (idle=182W) # 406 W with --vm 64 or 32 singularity exec /global/scratch/tin/singularity-repo/perf_tools_latest.sif stress --io 6 --hdd 2 --vm 64 -t 120 ## 25 Amp on n0098.savio2 (idle=6A/30W) singularity exec /global/scratch/tin/singularity-repo/perf_tools_latest.sif stress --io 6 --hdd 2 --vm 20 -t 120s ## 25 Amp on n0098.savio2 # 25A * 12V = 300W. 6A * 5V = 30W # -t 120 or -t 120s didn't stop process on savio1 :/ # process linger even after ^C . kill -9 dont work either, had to reboot!
stress-ng --cpu 4 --all # --all = parallel run of all tests
T=60; singularity exec /global/scratch/tin/singularity-repo/perf_tools_latest.sif stress-ng --cpu 63 --backoff 15 --timeout ${T} --tz --log-brief # 176W # don't see hdd access
T=60; singularity exec /global/scratch/tin/singularity-repo/perf_tools_latest.sif stress-ng --all 1 --backoff 15 --timeout ${T} --tz --log-brief # keeps crashing :(
cd /local/scratch/tin/tmp
stress-ng --all 1 --backoff 15 --timeout 120s --tz --log-brief
# --all 1 : all tests, 1 worker (enough to cause crazy high uptime)
rm /local/scratch/tin/tmp/stress*
stress-ng is newer, but didn't work on the knl node (at least not --all 1)
or did not use more power than stress test, so didn't bother.
maybe --vm would use more power... i dont need the stats from it
old stress dont leave files behind needing clean up ^_^
as benchmark comparing diff cpu,
stress-ng with specific cpu test can output bogo ops/sec metric:
running as sudo get access to more counters, eg cache, branch miss
singularity exec /global/scratch/tin/singularity-repo/perf_tools_latest.sif /bin/stress-ng --cpu-method matrixprod --metrics-brief --perf -t 100 --cpu 40
stressor bogo ops real time usr time sys time bogo ops/s bogo ops/s
(secs) (secs) (secs) (real time) (usr+sys time)
-------- -------- -------- -------- -------- ^^^^^^^^^^^ --------------
cpu 883734 300.01 9587.01 0.01 2945.73 92.18 # amd 7302P 1x16 with hyperthreading (thus --cpu 32) 210sav3
cpu 297171 100.00 3197.30 0.00 2971.57 92.94 # amd 7302P 1x16 with hyperthreading ran as --cpu 32 -t 100
cpu 292047 100.00 1599.19 0.00 2920.40 182.62 # amd 7302P 1x16 with hyperthreading ran as --cpu 16 -t 100
cpu 1216396 300.00 5996.84 0.00 4054.61 202.84 # intel E5-2670 v2 @ 2.5 GHz 2x10 hyperthreading off (ran as --cpu 32) 152sav1 ivy bridge
cpu 1384328 300.03 5996.58 0.13 4613.97 230.85 # intel E5-2670 v2 @ 2.5 GHz 2x10 hyperthreading off (ran as --cpu 20) 152sav1
cpu 1555260 300.00 11965.20 0.05 5184.13 129.98 # intel 6230 cascadelake 2x20 2.1GHz --cpu 40 192sav3
cpu 520460 100.00 3989.61 0.00 5204.38 130.45 # intel 6230 -t 100, so that bogo ops is mult of ops/s
7z apt install p7zip-full # 7z yum install p7zip # 7za 7z b --mmt1 # multi-threaded benchmark via 7z compression alg, report MIPS. singularity exec /global/scratch/tin/singularity-repo/perf_tools_latest.sif /usr/bin/7za b # need to loop it for stress test, as (b)enchmark finishes quickly # there are some idle time, Power ranged 144-160W # so not as good of power load test as "stress" mprime no ready yum package finding prime was said to be good exercise for cpu/ram but many sub-tasks, benchmark , need some effort to do benchmark result comparison. xref CF_BK/sw/mprime.rst singularity exec /global/scratch/tin/singularity-repo/perf_tools_latest.sif /bin/sysbench --threads=32 cpu --cpu-max-prime=20000 run sysbench --threads=32 cpu run output is events/second for cpu speed, not sure what to make of it.For single machine with display, consider
hashcat is a password cracker, actually it claim to be fastest. has a benchmark option and support cpu, GPU OpenCL runtime from even AMD/ROCm.
traceroute # some use icmp to trace, other use udp. traceroute --port=623 172.10.0.128 # potentially check rachability of ipmi udp port traceroute --port=33434 is the default, it is Destination UDP port that this pokes at. If host is reachable, even if destination udp port is not open/listen mode, traceroute will work.
tcpdump parameters -n: ip number, do no resolve hostname -e: ethernet (?) -i: interface -s 16000 : set capture frame size to 16k -w [FILE] : write output to file (capture use, more info than redirect output) file can be opened by wireshark -r [FILE] : read from to file , use -vv to add verbosity to parsed output host IP-or-NAME : capture info only related to the specified host operators accepted: && = and || = or ! = not eg cmd of tcpdump [expression] : tcpdump host 10.0.71.165 # bi-directional traffic tcpdump src 10.0.71.165 -w myfile.tcpdump # -w write output to file tcpdump 'dst net 128.3' tcpdump 'src or dst port ftp-data' tcpdump 'ether host 0:d0:b7:a9:c9:5a' tcpdump -v -i enp0s3 ether host 3c:ec:ef:bb:b5:bb or 90:e2:ba:c5:53:00 # verified will capture traffic for both mac tcpdump icmp -w myfile.tcpdump # only icmp traffic, all hosts tcpdump -v -v host 10.0.71.165 # -v -v, increase verbosity. # usable in screen output for low amount of traffic tcpdump -vv -e -n -i enp0s3 port 67 or port 68 # these ports are arp req (bootps) and reply (bootpc) used by DHCP tcpdump host 10.0.71.165 -i eth2 -s 64000 -w myfile.tcpdump # -i eth2, only traffic flowing thru specified nic # -s 64000, max packet capture size. 64k is probably bigger than needed, but works well # -s 0 would have no limit on caputre size, ie capture as large as needed # -w write output to file tcpdump -vv -i eno1 'ether host f8:f2:1e:66:81:50 || ether host ac:1f:6b:bb:63:00' tcpdump -i eno1 'ether host f8:f2:1e:66:81:50 || ether host ac:1f:6b:bb:63:00' tcpdump -e -i eno1 'ether host f8:f2:1e:66:81:50 || ether host ac:1f:6b:bb:63:00' # -e would capture eth frame instead of ip packet ? sudo tcpdump -i eno1 'dst 10.22.4.200 || dst 10.22.22.200 || dst 10.22.44.200' sudo tcpdump -v -n host 10.22.22.55 -i enp94s0f0 tcpdump -u port 69 -i enp94s0f0 # tftp traffic on udp/69 sudo tcpdump -e -v -n ether host ac:9e:17:2f:95:c0 or ether host ac:9e:17:2f:95:bf -i enp94s0f0 # -e add mac address to output sudo tcpdump port 514 -i enp216s0f0 -S -X # hex dump of syslog traffic on port 514, let it capture both tcp and udp (@@ vs @ in rsyslog.conf tcpdump ip6 -i enp0s8 # monitor IPv6 traffic only tcpdump -i enp0s8 ip6 and port 25 # IPv6 and mail port wireshark display filter tcp.port==25 ref: https://wiki.wireshark.org/DisplayFilters
showmount -e 192.168.209.30 # VIP
tcpdump -n host 172.24.51.182 # misconfigured NAT
18:49:41.964873 eth0 < 172.24.51.182 > tin-linux.zambeel.com: icmp: 172.24.51.182 udp port sunrpc unreachable [tos 0xc0]
18:56:24.677264 eth0 < 172.24.51.182 > 10.0.15.11: icmp: 172.24.51.182 udp port sunrpc unreachable [tos 0xc0]
18:56:24.679401 eth0 < 172.24.51.182 > 10.0.15.11: icmp: 172.24.51.182 udp port sunrpc unreachable [tos 0xc0]
timestamp src-if ? source ip destination prtl err message
tcpdump -n port sunrpc
18:54:31.055821 eth0 > 10.0.15.11.1388 > 192.168.209.30.sunrpc: udp 56
src-if ? source ip.port ? dest ip.port : protocol + port
[z-00D0B7A873CE] # tcpdump -e port sunrpc
18:15:55.628675 eth2 < 0:e0:52:d:7e:18 0:0:0:0:0:1 ip 74: 10.0.15.11.2499 > 172.24.51.182.sunrpc: S 4260207884:4260207884(0) win 32120 (DF)
time if ? src mac dst-mac(host) src ip.port dest ip.port TCP SYN and other protocol info
18:15:55.628696 eth2 > 0:0:0:0:0:0 0:2:e3:0:3b:9d ip 54: 172.24.51.182.sunrpc > 10.0.15.11.2499: R 0:0(0) ack 4260207885 win 0
time if ? src mac dst-mac(host) src ip.port dest ip.port TCP SYN and other protocol info
Here is an example of messed up translation.
Note that source & dest mac-address is rewritten on each router hop.
[z-00D0B7A871DF] # tcpdump -n | egrep '10\.0\.15\.11|192\.168'
19:02:43.964206 eth2 > 172.24.51.12.telnet > 10.0.15.11.2411: P 2646085534:2646085754(220) ack 2623622447 win 32120 {nop,nop,timestamp 2624922 80719743} (DF)
19:02:43.982115 eth2 < 10.0.15.11.2411 > 172.24.51.12.telnet: . 1:1(0) ack 220 win 31856 {nop,nop,timestamp 80720053 2624922} (DF)
19:02:45.277592 eth2 B 172.24.51.1.route > 172.24.51.255.route: rip-resp 25: {192.168.13.0/255.255.255.0}(2) {192.168.14.0/255.255.255.0}(2) {192.168.15.0/255.255.255.0}(2) {192.168.16.0/255.255.255.0}(2) {192.168.17.0/255.255.255.0}(2)[|rip]
snoop host [IP] # traffic with a given host (as src or dst)
snoop -r port 25 # all traffic in port 25 (smtp),
# do not resolve ip to dns names
-s = sniplet length (def is whole packet)
= 80 ip hdr only, 120 = nfs header only
-V = layer info
-v = more verbose than -V, lot of info.
from cli :
Usage: snoop
[ -a ] # Listen to packets on audio
[ -d device ] # settable to le?, ie?, bf?, tr?
[ -s snaplen ] # Truncate packets
[ -c count ] # Quit after count packets
[ -P ] # Turn OFF promiscuous mode
[ -D ] # Report dropped packets
[ -S ] # Report packet size
[ -i file ] # Read previously captured packets
[ -o file ] # Capture packets in file
[ -n file ] # Load addr-to-name table from file
[ -N ] # Create addr-to-name table
[ -t r|a|d ] # Time: Relative, Absolute or Delta
[ -v ] # Verbose packet display
[ -V ] # Show all summary lines
[ -p first[,last] ] # Select packet(s) to display
[ -x offset[,length] ] # Hex dump from offset for length
[ -C ] # Print packet filter code
Capture traffic on NIC hme0 specific to a host, capture up 8K of the packet, and dump result to an output file: snoop -d hme0 -s 8192 -o /tmp/snoop.out host 10.215.55.211 Read input file back. May wish to use ethereal to read this file for easier access. snoop -i /tmp/snoop.out snoop -s 120 port 25 host 211.196.53.194 titaniumleg.com mail server traffic monitor snoop -r -D -P -s 1500 -c 100000 -o /export/tmp/smtp01.20030122.snoop port 25 snoop -n /dev/null -D -P -s 1500 -c 100000 -o /export/tmp/smtp01.20030122.snoop port 25 snoop -D -s 9000 -c 100000 -o jumpstartclient.snoop host jumpstartclient -r = do not resolve hostname # not in sol 7 snoop -D = display num of dropped packets -P = non promiscuous mode capture (don't use in troubleshooting jumpstart problems). -s snipplet length -c count num of backets to capture -o output file ### ### more explanations TBA ###
ethereal / wireshark (GUI)
tethereal / tshark (CLI)
most flags work for both, but not for capture filters!.
snoop-like behaviour (mostly for ethereal):
-l : scroll capture
-S : update as capture is in progress.
-k : start capture immediately (disable interaction?)
-f : specify a capture Filter
-i [IF] : specify interface, eg eth0, hme0
-n : no dns resolution, use ip Number
-V : more verbose output, captured data displayed in tree mode instead of 1 line per packet.
-f : capture filter expression (tcpdump notation needed), eg:
tcp port 23 and host 10.0.0.5
src net 10.0.15.0/24
dst net 10.0.15.0 mask 255.255.255.0
[src|dst] host HOSTNAME
ether [src|dst] host 00:E0:2B:DE:0E:00
[tcp|udp] [src|dst] port PORTNUM
eg of multiple host:
host 10.215.20.152 || host 10.215.2.21 || host 10.215.19.73
tshark -l -S -k -i eth0 -f "host 12.34.65.123" # human readable live capture against a remote host using console
------------------------------------------------------------
ethereal view filter expression
[ work in GUI filter box when viewing,
NOT as capture filter (which is tcpdump format ]
operatos:
eq, == Equal
ne, != Not equal
gt, > Greater than
lt, < Less Than
ge, >= Greater than or Equal to
le, <= Less than or Equal to
and, && Logical AND
or, || Logical OR
not, ! Logical NOT
boolean: true (1) or false (0)
some commonly used filter fields:
eth.src == aa-aa-aa-aa-aa-aa
ip.dst eq www.mit.edu
ip.src == 192.168.1.1
ip.addr == 129.111.0.0/16
eth.src == aa-aa-aa-aa-aa-aa
eth.src[0:3] == 00:00:83 # filter by vendor by use of slide
tcp.port == 80 and ip.src == 192.168.2.1
ip.addr is for both src or dest, these multiple ocurring field is a bit confusing for packet filtering.
ip.addr == 10.243.197.118 && tftp.opcode == 1 # tftp.opcode is Read Request (ie, the packet that ask for file)
for generic filter dealing with a specific host, but not necessary filtering by tcp/udp/icmp
ip.dst
ip.src
ip.addr
udp
udp.port
udp.dstport
udp.srcport
tcp
tcp.port
tcp.dstport
tcp.srcport
tcp.seq
icmp
bootp.dhcp==true : frame is dhcp
bootp.hw.addr
smb.cmd==(unsigned 8 bit int) : smb protocol command number
smb.cmd == 0x06 : cmd is smb unlink
smb.status != 0x0000 : Error code, 4 bytes aka status, lot of items.
smb.errcls != 0x0 : error class, 1 byte represent the categories
0x0 = Success
0x1 = DOS Error
0x2 = Server Error
0x3 = hardware error
0x4 = not a smb cmd
Note, netBench Fail code 32 maybe in Dos or Hrd.
smb.pid
smb.mid (multiplex id)
smb.uid (user id, maybe per process)
nfs.*
nfs.fh.version != 3 = not sure what this is, not nfs protocol version!
rpc.programversion != 3 = all packet that are rpc program nfs version 3.
lot of higher level protocol stuff available, including vlan on switches, etc.
see the man page on ethereal or tethereal (very long!)
GUI version, filter can just enter a protocol type. eg: smb
That means smb protocol is present. A protocol in the filter w/o any comparison operator means filter packets where such field is present in the packet.
eg: smb.errcls filter packet that contain smb error class.
Network trace capture with tcpdump or snoop, save to file for viewing with ethereal
tcpdump -i [interface] -s 1500 -w [some-file]
tcpdump -s 8192 -w netuse.tcpdump 'host 10.0.71.232 or host 10.0.71.15'
snoop -d hme0 -o /tmp/snoop.out host 10.215.55.211
editcap can be used to trim captured file, or convert between formats
(tcpdump, ethereal, snoop, ms netmon, etc).
fe80:: # ipv6 link local, autogenerated (like IPv5 169.254...) can't go thru router, but usable in same subnet
fe80::ec4:7aff:fe75:c9b0 # bofh ipv6 link local
0cc4:7a75:c9b0 # bofh mac
fe80::221:ccff:fed3:3d41 # cueball ipv6 link local
0021:ccd3:3d41 # cueball mac
ping6 fe80::ec4:7aff:fe75:c9b0%2 # works (same host or another host on same subnet). %2 is interface id as per "ip a"
https://askubuntu.com/questions/546405/ping-adress-ipv6
nmap: network scanner nmapfe: w/ gui front end, supposed to need gtk, but worked anyway. nmap -sT -O -PI -PT 172.27.31.0/24 # scan whole class C vlan 31, with os identification. long output. sudo nmap -p1-62000 somehost # scan port 1 to 62000 against machine named somehost sudo nmap -sU somehost -p8255 # check UDP port 8255 to see if it is open sudo nmap -e enp0s9 freeipa # check all TCP ports to IPA server is reachable sudo nmap -e enp0s9 -sU freeipa # check all UDP ports to IPA server is reachable, takes ~18 min, another took 33 min sudo nmap -e enp0s9 -sU -F freeipa # -sU = UDP scan. -F = fast mode, scan fewer ports than default, still takes long time sudo nmap -e enp0s9 -sU -p53-1813 freeipa # scan specified port range only sudo nmap -e enp0s9 -sUT -p53,123,80,88,443,8443,8080,389,464,636,1812-1813 freeipa # -sUT scan both UDP and TCP nmap -6 ::1 # scan localhost nmap -6 fe80::ec4:7aff:fe75:c9b0 # bofh ipv6 link local , can't go thru router, but should be usable in same "subnet" nmap -Pn -p8255 -6 fe80::ec4:7aff:fe75:c9b0 # scan works when on same host. nmap -Pn -p1-10000 -6 fe80::ec4:7aff:fe75:c9b0%2 # from remote, works after specifying which interface to use (%2, per ip a)